DeepMind

Artificial Intelligence beyond AlphaGo

DeepMind Technologies Limited is a British company now owned by Google. DeepMind is perhaps best known for AlphaGo, its AI-based software that plays the game of Go, evidently at least to a world championship standard. However, DeepMind's AI patenting activities indicate other interests going beyond board games and covering various fields of application of AI.

DeepMind Patent Application Families

Up to the present time1 , members of around 58 DeepMind families2 of patent applications have been published. Titles and application/patent numbers of key members3 of each family are shown in Table 4 below. A total of 85 members of the families have been published up to the present time1. Members of 4 families were published in 2014, having filing dates in 2012, 2013, and 2014. However, members of the other families are more recent, being published in 2018 and 2019, with application dates from 2016 to 2018. It is to be expected that more families and family members will appear as further publications take place.

DeepMind Patents

As would be expected with such a young portfolio, few patents have been granted so far. These are listed in Table 1. Claim 1 of each of the US patents is reproduced further below. So far, there are no granted European patents in the published record1.

 

Table 1

DeepMind Patents

CN105144203(B)

Signal processing systems

US9342781(B2)

 

US10032089(B2)

Spatial transformer

 

modules

US10176424(B2)

Generative neural

 

networks

US10198832(B2)

Generalizable medical

 

image analysis using

 

segmentation and

 

classification neural

 

networks

US8644607(B1)

Method and apparatus for

US8971669(B2)

image processing

 

Fields of Interest for DeepMind

Although DeepMind's interests go beyond board games, they appear to be focused on a narrow range of technologies, which is indicated by detailed classifications of the members of the DeepMind portfolio under the Cooperative Patent Classification (CPC) hierarchy1 .

Each patent or application is usually given more than one detailed classification to main groups or sub-groups of the CPC. DeepMind's families have been given 235 detailed classifications, all falling within only 9 CPC sub-classes, as shown in Table 2. The vast majority of detailed classifications fall within sub-class G06N, which covers – amongst other things – neural networks. Seen in more detail, the 235 classifications range across only 65 CPC main/sub-groups, with 80% of those classifications accounted for by 17 main/sub-groups as shown in Table 3. These focus on neural networks, machine learning, and the context of natural language processing, image analysis, speech synthesis etc., with a surprise mention of musical instruments.

   

Table 2

CPC sub-class

No. of detailed classifications

CPC description

 

within sub-class

 

G06N

187

Computer systems based on specific computational models

G06K

18

Recognition of data; presentation of data; record carriers;

   

handling record carriers

G06F

11

Electric digital data processing

G06T

8

Image data processing or generation, in general

G10L

5

Speech analysis or synthesis; speech recognition; speech or

   

voice processing; speech or audio coding or decoding

G10H

2

Electrophonic musical instruments

H04N

2

Pictorial communication, e.g. television

G05B

1

Control or regulating systems in general; functional elements

   

of such systems; monitoring or testing arrangements for such

   

systems or elements

Y04S

1

Systems integrating technologies related to power network

   

operation, communication or information technologies for

   

improving the electrical power generation, transmission,

   

distribution, management or usage, i.e. smart grids

   

Table 3

CPC main group/

No. of detailed classifications

CPC description

sub-group

to main group/sub-group

 

G06N3/0454

38

Computer systems based on biological models; Architectures,

   

e.g. interconnection topology; using a combination of multiple

   

neural nets

G06N3/08

31

Computer systems based on biological models; Learning

   

methods

G06N3/04

22

Computer systems based on biological models; Architectures,

   

e.g. interconnection topology

G06N3/0445

22

Computer systems based on biological models; Architectures,

   

e.g. interconnection topology; Feedback networks, e.g.

   

hopfield nets, associative networks

G06N3/006

17

Computer systems based on biological models; Physical

   

realisation, i.e. hardware implementation of neural networks,

   

neurons or parts of neurons

G06N3/084

13

Computer systems based on biological models; Learning

   

methods; Back-propagation

G06N3/088

10

Computer systems based on biological models; Learning

   

methods; Non-supervised learning, e.g. competitive learning

G06N3/0472

9

Computer systems based on biological models; Architectures,

   

e.g. interconnection topology; using probabilistic elements,

   

e.g. p-rams, stochastic processors

G06N3/00

8

Computer systems based on biological models

G06N3/063

3

Computer systems based on biological models; Physical

   

realisation, i.e. hardware implementation of neural networks,

   

neurons or parts of neurons

G06F17/2818

2

Digital computing or data processing equipment or methods,

   

specially adapted for specific functions; Processing or

   

translating of natural language; Statistical methods, e.g.

   

probability models

G06K9/4628

2

Methods or arrangements for reading or recognising printed

   

or written characters or for recognising patterns, e.g.

   

fingerprints; Extraction of features or characteristics of the

   

image; integrating the filters into a hierarchical structure

G06N3/02

2

Computer systems based on biological models; using neural

   

network models

G06N3/0481

2

Computer systems based on biological models; Architectures,

   

e.g. interconnection topology; Non-linear activation functions,

   

e.g. sigmoids, thresholds

G06N3/082

2

Computer systems based on biological models; Learning

   

methods; modifying the architecture, e.g. adding or deleting

   

nodes or connections, pruning

G10H2250/311

2

Aspects of algorithms or signal processing methods without

   

intrinsic musical character, yet specifically adapted for or used

   

in electrophonic musical processing; Neural networks for

   

electrophonic musical instruments or musical processing, e.g.

   

for musical recognition or control, automatic composition or

   

improvisation

G10L13/00

2

Speech synthesis; Text to speech systems

DeepMind Patent US9342781 (B2)

  1. A neural network system implemented as one or more computers for generating samples of a particular sample type, wherein each generated sample belongs to a respective category of a predetermined set of categories, and wherein each generated sample is an ordered collection of values, each value having

a sample position in the collection, and wherein the system comprises:

a first stochastic layer configured to stochastically select a category from the predetermined set of categories;

a first deterministic subnetwork configured to: receive an embedding vector corresponding to the selected category, and

process the embedding vector to generate a respective sample score for each sample position in the collection; and

a second stochastic layer configured to generate an output sample by stochastically selecting, for each sample position, a sample value using the sample score for the sample position.

DeepMind Patent US10032089 (B2)

  1. An image processing neural network system implemented by one or more computers, wherein the image processing neural network system is configured to receive one or more input images and to process the one or more input images to generate a neural network output from the one or more input images, the image processing neural network system comprising:

a spatial transformer module, wherein the spatial transformer module is configured to perform operations comprising:

receiving an input feature map derived from the one or more input images, and

applying a spatial transformation to the input feature map to generate a transformed feature map, comprising:

processing the input feature map to generate, based on the input feature map, spatial transformation parameters that define the spatial transformation to be applied to the input feature map, and

sampling from the input feature map in accordance with the spatial transformation parameters generated based on the input feature map to generate the transformed feature map.

DeepMind Patent US10176424 (B2)

  1. A neural network system implemented by one or more computers, the neural network system comprising:

a recurrent neural network that is configured to, for each time step of a predetermined number of time steps, receiv e a set of latent variables for the time step and process the set of latent variables to update a hidden state of the recurrent neural network; and

a generative subsystem that is configured to:

for each time step of the predetermined number of time steps:

generate the set of latent variables for the time step and provide the set of latent variables as input to the recurrent neural network;

update a hidden canvas using the updated hidden state of the recurrent neural network; and

for a last time step of the predetermined number of time steps:

generate an output image using the updated hidden canvas for the last time step.

DeepMind Patent US10198832 (B2)

  1. A system comprising one or more computers and one or more storage devices storing instructions that when executed by the one or more computers cause the one or more computers to implement:

a first set of one or more segmentation neural networks, wherein each segmentation neural network in the first set is configured to:

receive an input image of eye tissue captured using a first imaging modality; and

process the input image to generate a segmentation map that segments the eye tissue in the input image into a plurality of tissue types;

a set of one or more classification neural networks, wherein each classification neural network is configured to:

receive a classification input derived from a segmentation map of eye tissue; and

process the classification input to generate a classification output that characterizes the eye tissue; and a subsystem configured to:

receive a first image of eye tissue captured using the first imaging modality;

provide the first image as input to each of the segmentation neural networks in the first set to obtain one or more segmentation maps of the eye tissue in the first image;

generate, from each of the segmentation maps, a respective classification input; and

provide, for each of the segmentation maps, the classification input for the segmentation map as input to each of the classification neural networks to obtain, for each segmentation map, a respective classification output from each classification neural network; and

generate, from the respective classification outputs for each of the segmentation maps, a final classification output for the first image.

DeepMind Patent US8644607 (B1)

  1. A method for processing an image to generate a signature which is characteristic of a pattern within said image comprising:

receiving an image;

overlaying a window at multiple locations on said image to define a plurality of sub-images within said image, with each sub-image each having a plurality of pixels having a luminance level;

determining a luminance value for each said sub-image, wherein said luminance value is derived from said luminance levels of said plurality of pixels;

combining said luminance values for each of said sub-images to form said signature;

wherein said combining is such that said signature is independent of the location of each sub-image.

DeepMind Patent US8971669 (B2)

  1. A non-transitory computer readable medium storing a computer program code that, when executed by one or more computers, causes the one or more computers to perform operations for processing an image to generate a signature which is characteristic of a pattern within the image, the operations comprising:

receiving an image;

overlaying a window at multiple locations on the image to define a plurality of sub-images within the image, with sub-image having a plurality of pixels having a luminance level;

determining a luminance value for each sub-image, wherein said luminance value is derived from the luminance levels of the plurality of pixels in the sub-image; and

combining the luminance values for each of the sub-images to form a signature for the image;

wherein the combining is such that the signature is independent of the location of each sub-image.

 

Table 4

Title

Publication number

Method And Apparatus For Image Searching

US2014019484 (A1)

Method And Apparatus For Conducting A Search

US2014019431 (A1)

Method And Apparatus For Image Processing

US2014185959 (A1);

 

US8971669 (B2)

Signal Processing Systems

GB2513105 (A)

Generative Neural Networks

US10176424 (B2);

 

US2017228633 (A1)

Umgebungsnavigation Unter Verwendung Von Verstärkungslernen

DE202017106697 (U1)

Processing Sequences Using Convolutional Neural Networks

WO2018048945 (A1)

Generating Video Frames Using Neural Networks

WO2018064591 (A1)

Neural Networks For Selecting Actions To Be Performed By A Robotic Agent

WO2018071392 (A1)

Reinforcement Learning With Auxiliary Tasks

WO2018083671 (A1)

Sequence Transduction Neural Networks

WO2018083670 (A1)

Recurrent Neural Networks

WO2018083669 (A1)

Scene Understanding And Generation Using Neural Networks

WO2018083668 (A1)

Reinforcement Learning Systems

WO2018083667 (A1)

Training Action Selection Neural Networks

WO2018083532 (A1)

Continuous Control With Deep Reinforcement Learning

MX2018000942 (A)

Data-Efficient Reinforcement Learning For Continuous Control Tasks

WO2018142212 (A1)

Memory Augmented Generative Temporal Models

WO2018142378 (A1)

Neural Programming

EP3360082 (A1)

Augmenting Neural Networks With External Memory

KR20180091850 (A)

Neural Episodic Control

WO2018154100 (A1)

Multiscale Image Generation

WO2018154092 (A1)

Action Selection For Reinforcement Learning Using Neural Networks

WO2018153807 (A1)

Training Machine Learning Models

WO2018153806 (A1)

Dueling Deep Neural Networks

US2018260689 (A1)

Asynchronous Deep Reinforcement Learning

US2018260708 (A1)

Training Neural Networks Using Posterior Sharpening

WO2018172513 (A1)

Selecting Action Slates Using Reinforcement Learning

EP3384435 (A1)

Distributional Reinforcement Learning

WO2018189404 (A1)

Black-Box Optimization Using Neural Networks

WO2018189279 (A1)

Generating Images Using Neural Networks

CN108701249 (A)

Training Neural Networks Using A Prioritized Experience Memory

CN108701252 (A)

Training Neural Networks Using Normalized Target Outputs

CN108701253 (A)

Associative Long Short-Term Memory Neural Network Layers

EP3398118 (A1)

Compressing Images Using Neural Networks

EP3398114 (A1)

Augmenting Neural Networks With External Memory

EP3398117 (A1)

Generating Audio Using Neural Networks

US2018322891 (A1)

Processing Text Sequences Using Neural Networks

US2018329897 (A1)

Spatial Transformer Modules

US2018330185 (A1)

Generating Output Examples Using Bit Blocks

US2018336455 (A1)

Programmable Reinforcement Learning Systems

WO2018211146 (A1)

Making Object-Level Predictions Of The Future State Of A Physical System

WO2018211144 (A1)

Neural Network System

WO2018211143 (A1)

Imagination-Based Agent Neural Networks

WO2018211142 (A1)

Imagination-Based Agent Neural Networks

WO2018211141 (A1)

Data Efficient Imitation Of Diverse Behaviors

WO2018211140 (A1)

Training Action Selection Neural Networks Using A Differentiable Credit Function

WO2018211139 (A1)

Multitask Neural Network Systems

WO2018211138 (A1)

Neural Network Systems For Action Recognition In Videos

WO2018210796 (A1)

Training Action Selection Neural Networks Using Look-Ahead Search

WO2018215665 (A1)

Noisy Neural Network Layers

WO2018215344 (A1)

Training Action Selection Neural Networks

WO2018224695 (A1)

Generating Discrete Latent Representations Of Input Data Items

WO2018224690 (A1)

Selecting Actions Using Multi-Modal Inputs

WO2018224471 (A1)

Feedforward Generative Neural Networks

US2018365554 (A1)

Generalizable Medical Image Analysis Using Segmentation And Classification

US10198832 (B2);

Neural Networks

US2019005684 (A1)

Training Action Selection Neural Networks Using Apprenticeship

WO2019002465 (A1)

Learning Visual Concepts Using Neural Networks

WO2019011968 (A1)

The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.